How ChatGPT, Perplexity and Google AI Overviews each pick their sources

“Optimise for AI” is advice that falls apart the moment you ask “which AI?” Google's AI Overviews, ChatGPT and Perplexity all answer questions and cite sources, but they assemble those answers in meaningfully different ways — and a page tuned for one can be invisible in another. If you want to be the source these systems name, it helps to understand how each one actually decides who gets quoted. We have spent a year watching all three, and the differences are consistent enough to plan around.

The short version: AI Overviews reward sites that already rank and carry conventional authority; Perplexity rewards the clearest, most directly-useful page it can retrieve in the moment; and ChatGPT sits between them, leaning on what it already knows about an entity but increasingly willing to browse and cite fresh sources. None of them rewards tricks. All of them reward being genuinely useful and genuinely trusted — they just weight those two things differently. Here is how each one works, and what that means for your pages.

Google AI Overviews: authority first, structure second

AI Overviews are built directly on top of Google's existing search index, and that single fact explains most of their behaviour. The candidate pool is, roughly, pages that already rank for the query and its sub-questions. Google retrieves from that pool, extracts passages, and composes the overview with links back to the sources it leaned on. The practical consequence is blunt: if you do not have the conventional authority to rank on page one, you are rarely in the running for the overview, because you were never in the candidate set. This is why classic SEO has not gone away — for AI Overviews it is the entry ticket. Once you are in the pool, structure decides whether you get the pull-quote: answer-first sections, clear headings, and FAQ markup make your passage the easy one to lift. So the AI Overview playbook is two-layered: earn the authority to rank at all, then structure the page so your passage is the cleanest answer to the specific question.

Perplexity: the clearest answer in the moment wins

Perplexity behaves less like a search engine with a summary bolted on and more like a researcher who reads a handful of pages for each question and writes up what they found, footnotes and all. It runs fresh retrieval at query time, weighs how directly each page answers the exact question, and cites the ones that answered it best. The important difference from AI Overviews is that Perplexity is noticeably more willing to cite a smaller or newer site, provided that site's page is genuinely the clearest match. Conventional authority still helps — it influences what gets retrieved and trusted — but a focused, well-written page that nails a specific question can earn a citation here even without a huge backlink profile. That makes Perplexity the most accessible of the three for a smaller brand: the lever is page quality and directness, not just domain strength. If you write one genuinely excellent, self-contained answer to a real question, Perplexity is the engine most likely to reward it quickly.

ChatGPT: what it already knows, plus what it browses

ChatGPT is really two behaviours in one product. Without browsing, it answers from what it learned in training, which means it names the entities, brands and sources it saw often and consistently described — being well-known and well-described is everything in that mode, and there is no page you can publish today that changes a weight set months ago. With browsing enabled, it behaves more like Perplexity: it retrieves live, reads, and cites current sources. The strategic read is that ChatGPT rewards two different things on two different timescales. The slow game is becoming an entity ChatGPT knows from the wider web — consistent mentions, a clear identity, corroboration across trusted sources — so that even its no-browsing answers tend to name you. The fast game is having pages clear and current enough that when it does browse, yours is the one it quotes. Both games reward the same underlying thing: a brand the web talks about clearly, with pages that answer questions cleanly.

What all three agree on

For all their differences, the three engines converge on the same two preferences, and that convergence is the thing to build around. First, they all favour pages that answer a specific question directly and self-sufficiently — a passage that stands alone, leads with the answer, and does not make the model wade through preamble. Second, they all favour sources they have reason to trust, whether that trust comes from ranking authority, from fresh relevance signals, or from an entity being well-established across the web. A page that is both a clean, direct answer and a trusted source is the one that gets cited everywhere. That is not three strategies; it is one strategy that happens to pay off in three places. The engine-specific tactics are refinements on top of it, not replacements for it.

So what should you actually optimise?

Do not build three separate content strategies — build one, aimed at the shared preference, and tune at the margins. Start by earning genuine authority and a clear entity, because that is the foundation AI Overviews gate on and the thing that makes ChatGPT name you unprompted. Then write pages that answer specific questions directly, with answer-first structure and FAQ schema, because that is what wins the pull-quote in Overviews and the citation in Perplexity and browsing ChatGPT alike. Where you have the budget to tune further: for AI Overviews, invest in the conventional ranking strength that gets you into the candidate pool; for Perplexity, invest in the sharpest possible answer to each specific question, since directness can beat domain size there; for ChatGPT, invest in being described consistently across the web so its memory of you is accurate. But notice that every one of those refinements is a dialling-up of the same two levers — be trusted, be clear — not a different game.

How to know if it is working

The only way to manage this is to measure it, because guessing which engine names you is a waste of good strategy. Pick the questions a buyer in your category would ask, and ask them across the engines on a schedule — not once, but regularly, because these systems change and so do your competitors. Record, for each engine, whether you are named, who is named instead, and which page got cited. The patterns tell you where to spend: if Perplexity cites you but AI Overviews never do, you have a directness-but-not-authority profile, and the fix is conventional ranking strength. If ChatGPT describes you vaguely, your entity is fuzzy and the fix is corroboration. We built OptiLink's AI-visibility check to run exactly this loop across engines, so the next move is informed by what the machines are actually doing rather than by the latest confident thread about what they supposedly reward.

Retrieval versus memory

To understand source selection you have to understand that these engines work in two different modes, often at once. Some of what a model “knows” is baked into its parameters from training, and some is pulled in live through retrieval at the moment you ask. The distinction matters enormously for visibility. Content that only lives in training data is frozen at the cutoff and cannot reflect anything recent, while content that can be retrieved live is eligible to appear in answers about current topics the moment it is published. The practical upshot is that being retrievable — crawlable, fresh, clearly on-topic — is how you stay eligible for the answers that matter most, because the engines increasingly lean on retrieval for anything where currency counts, which is most commercial questions.

Freshness is weighted, but not evenly

Recency is a real factor in selection, but it is applied with judgement rather than as a blunt rule. For questions where the answer genuinely changes over time — prices, availability, current best options, anything tied to a moving field — the engines lean hard toward fresh sources and will pass over stale ones however authoritative. For questions where the answer is stable — definitions, established principles, how something fundamentally works — freshness matters far less and durable authority dominates. Knowing which kind of question you are competing on tells you how much to invest in keeping a page current: constant maintenance for the time-sensitive topics, and durable depth for the evergreen ones. Treating every page the same wastes effort on the stable ones and loses the volatile ones.

Domain trust versus page trust

Selection operates at two levels that reinforce each other. There is the trust the engines place in a domain as a whole — its track record, its recognition, its overall credibility — and the trust they place in a specific page for a specific question. A strong domain gives every page on it a head start, which is why establishing site-wide credibility pays off across all your content. But a strong domain does not rescue a weak page: on a specific question, a thorough, well-structured, clearly-relevant page will be chosen over a thin one even from a respected site. The two levels work together, and the winning position is to be a credible domain publishing genuinely strong individual pages, rather than relying on either alone to carry the other.

How engines handle conflicting sources

One of the harder problems these engines face is disagreement between sources, and how they resolve it shapes who gets cited. Faced with conflicting claims, the engines tend to favour the sources that are most corroborated, most clearly reasoned, and most credibly attributed — the claim that several trustworthy sources agree on, stated cleanly with evidence, beats a lone contrarian assertion however confident. This rewards being right in a way that aligns with the wider evidence rather than being provocatively different for its own sake. It also rewards showing your reasoning and sourcing, because a claim that explains and supports itself is easier for the engine to trust when it has to adjudicate. The lesson is to be the well-supported consensus-aligned source on contested points, not the loudest outlier.

Structured data as a selection aid

Structured data does not force an engine to cite you, but it materially helps selection by removing ambiguity. Schema that marks up what a page is, who wrote it, what question it answers, and how its parts relate makes it easier for an engine to determine, quickly and confidently, that your page is a relevant and credible answer to the query in front of it. In a selection process that is effectively a fast triage across many candidate sources, anything that makes your relevance and credibility unambiguous improves your odds. This is why FAQ schema, article schema, and clear entity markup repay the effort — not as a trick, but as a way of making the genuine relevance you already have legible to a machine deciding under time pressure.

The single-passage advantage

A subtle but decisive factor is whether an engine can find, on your page, a single self-contained passage that answers the question cleanly. These systems often need to lift a specific chunk, not synthesise your whole page, so a page that states the answer in one clear, complete, quotable passage is far easier to cite than one where the answer is diffused across several paragraphs that only make sense together. This is why answer-first writing wins: lead each section with the complete claim, make it stand on its own, and let the elaboration follow. You are not just being readable for humans; you are making sure that when an engine goes looking for the one passage it can quote and attribute, yours is the one that is ready to be lifted.

Being in the conversation off-site

Selection is not decided only by what is on your own pages. The engines read the wider web's treatment of you — whether other credible sources mention you, discuss you, and corroborate your expertise — and they use that as evidence of whether you are a source worth citing. This means part of earning selection happens entirely off your site, in the places your topic is discussed: being referenced in industry writing, named in comparisons, present in the forums and communities where your subject lives. A brand that the wider web treats as a reference is a safer bet for an engine than one that only asserts its own relevance, and that external corroboration is often the tie-breaker when several on-page-strong candidates compete.

What you can and cannot control

It helps to be clear-eyed about the limits. You cannot control the engines' ranking logic, their training data, or exactly which source they pick on a given day, and anyone promising deterministic placement is selling something. What you can control is your eligibility and your odds: whether your content is retrievable and fresh, whether it answers questions cleanly and quotably, whether your credibility is legible on and off the page, and whether you are genuinely one of the best answers to the questions you care about. The right posture is to maximise the things within your control and treat the rest as probability rather than a lever. Do the controllable work well and consistently, and selection stops being a mystery and becomes a matter of odds you have stacked in your favour.

Putting it together

None of these factors operates in isolation; selection is the product of all of them at once. The source that gets cited tends to be retrievable and current where currency matters, backed by both domain and page trust, aligned with the corroborated evidence, marked up clearly, written so a single passage can be lifted, and talked about off-site as a genuine reference. That sounds like a lot, but it reduces to a simple idea: be genuinely one of the best, clearest, most credible answers to the question, and make that fact as easy as possible for a machine to verify quickly. Everything in this article is a facet of that one principle, and teams that internalise it stop chasing individual tactics and start building the underlying credibility that every selection factor rewards.

A note on paid versus organic

One question that comes up often is whether paid placement influences AI source selection, and the honest answer is that for the organic, synthesised part of an AI answer, selection is driven by the credibility and relevance signals described throughout this article, not by ad spend. This is genuinely different from the paid dynamics of traditional search, and it has a liberating implication: you cannot buy your way into being the cited source, which means a smaller brand with genuinely superior, more credible content on a specific question can be named ahead of a larger competitor with a bigger budget. The engines are, in this respect, something of a meritocracy for substance. That does not mean advertising has no role in your wider strategy, but it does mean the work of earning citations is won on quality and credibility rather than spend — which is good news for anyone willing to do the harder, more durable work.

Frequently asked questions

Do ChatGPT, Perplexity and Google AI Overviews cite sources the same way?

No. AI Overviews favour pages that already rank and carry conventional authority; Perplexity runs fresh retrieval and cites the clearest direct answer, even from smaller sites; ChatGPT leans on entities it already knows and browses to cite fresh sources when asked.

Which AI engine is easiest for a smaller brand to get cited in?

Usually Perplexity. Because it retrieves live and rewards the most direct, self-contained answer to a specific question, a focused, well-written page can earn a citation there without a large backlink profile — directness can beat domain size.

Why does classic SEO still matter for AI Overviews?

Because AI Overviews are built on Google's index and mostly draw from pages that already rank for the query. If you lack the authority to rank on page one, you are rarely in the candidate pool, so conventional SEO is the entry ticket.

How do I get named by ChatGPT when it isn't browsing?

Become an entity it learned from the wider web — consistent mentions, a clear identity, and corroboration across trusted sources. No page published today changes weights set during training, so the lever is long-run reputation, not a single article.

Should I build a separate content strategy for each AI engine?

No. All three reward the same two things — a clear, self-contained answer and a trusted source — so build one strategy around that and tune at the margins. Measuring which engine names you (e.g. with OptiLink's AI-visibility check) tells you where to tune.